Computer Networks 57 (2013) 622–633

Contents lists available at SciVerse ScienceDirect

Computer Networks

journal homepage: www.elsevier .com/locate /comnet

A cross-layer communication module for the Internet of Things

Chong Han a,⇑, Josep Miquel Jornet a, Etimad Fadel b, Ian F. Akyildiz a,b

a Broadband Wireless Networking (BWN) Laboratory, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USAb Department of Information Technology, King Abdulaziz University, Jeddah, Saudi Arabia

a r t i c l e i n f o a b s t r a c t

Article history:Received 28 September 2012Accepted 5 October 2012Available online 22 October 2012

Keywords:Internet of ThingsCyber-physical systemsCross-layer designOptimization

1389-1286/$ - see front matter � 2012 Elsevier B.Vhttp://dx.doi.org/10.1016/j.comnet.2012.10.003

⇑ Corresponding author. Tel.: +1 404 894 6616.E-mail addresses: chong.han@ece.gatech.edu (C.

gatech.edu (J.M. Jornet), eafadel@kau.edu.sa (E. Faedu (I.F. Akyildiz).

The Internet of Things (IoT) is a novel networking paradigm which allows the communica-tion among all sorts of physical objects over the Internet. The IoT defines a world-widecyber-physical system with a plethora of applications in the fields of domotics, e-health,goods monitoring and logistics, among others. The use of cross-layer communicationschemes to provide adaptive solutions for the IoT is motivated by the high heterogeneityin the hardware capabilities and the communication requirements among things. In thispaper, a novel cross-layer module for the IoT is proposed to accurately capture both thehigh heterogeneity of the IoT and the impact of the Internet as part of the network archi-tecture. The fundamental part of the module is a mathematical framework, which is devel-oped to obtain the optimal routing paths and the communication parameters amongthings, by exploiting the interrelations among different layer functionalities in the IoT.Moreover, a cross-layer communication protocol is presented to implement this optimiza-tion framework in practical scenarios. The results show that the proposed solution canachieve a global communication optimum and outperforms existing layered solutions.The novel cross-layer module is a primary step towards providing efficient and reliableend-to-end communication in the IoT.

� 2012 Elsevier B.V. All rights reserved.

1. Introduction actuation functionalities into daily things and communi-

Nowadays, the Internet is used by more than two billioncustomers around the world to browse content, send andreceive emails, access multimedia resources, play onlinegames, and make social networking, among others. More-over, the Internet is also expected to serve as a global plat-form to interconnect physical objects or ‘‘things’’, thus,enabling new ways of working, interacting, entertaining,and living [1,2].

Within such perspective, the Internet of Things (IoT) is anovel networking paradigm which allows the communica-tion among all sorts of physical objects over the Internet[3,4]. The IoT is enabled by embedding communicationcapabilities and, in some cases, identification, sensing and

. All rights reserved.

Han), jmjornet@ece.del), ian@ece.gatech.

cating in extended Internet technologies. The IoT definesa truly world-wide cyber-physical system in which everyphysical object can be connected and controlled remotely.The ensemble of applications and services leveraging suchtechnologies open a plethora of new business and marketopportunities in the fields of domotics, e-health, real-timemonitoring of industrial processes, and intelligent logistics,among others [5,6].

A major bottleneck in the IoT is posed by the very highheterogeneity in both the hardware capabilities and thecommunication requirements among different types ofthings.

� From the hardware perspective, things can have verydifferent computation, memory, power or communica-tion capabilities. For instance, a cellular phone or a tab-let has much better communication and computationcapabilities than a single-purpose electronic productsuch as a heart rate monitor watch.

C. Han et al. / Computer Networks 57 (2013) 622–633 623

� Things can have very different Quality of Service (QoS)requirements in terms of delay, energy consumptionor reliability. For example, minimizing the energy forcommunication/computation purposes is a major con-straint for the batter-powered devices without efficientenergy harvesting techniques. On the contrary, thisenergy constraint is not critical for the devices withpower supply connection.

These two characteristics pose the conflict in designing aunifying framework which can take care of diversity ofcapabilities and functionalities of things. As a result,the heterogeneity traits of the network motivate theuse of adaptive cross-layer communication schemes forthe IoT.

There exists several cross-layer protocols for WirelessSensor Networks (WSNs) [7,8], Wireless Mesh Networks(WMNs) [9] or Ad Hoc Networks (AHNs) [10]. However,they cannot be applied to the IoT due to several reasons.First, the heterogeneity of the IoT incurs that things havelargely diverse hardware capabilities, different QoSrequirements and individual goals. On the contrary, inWSNs, nodes usually have very similar hardware specifica-tions, common communication requirements and a sharedgoal. Second, the Internet is involved in the IoT networkarchitecture, from which it inherits a centralized and hier-archical architecture. In comparison, in WSNs, WMNs andAHNs, highly flat network architectures are considered, inwhich nodes communicate in a multi-hop fashion and alsothe Internet is not involved.

In this paper, we propose a novel cross-layer optimiza-tion framework and communication module for the IoT.Our proposed mathematical framework captures the highdevice and service heterogeneities of the IoT. In particular,it exploits the interrelations among the device specifica-tions, physical layer, link layer and network layer, to findthe optimal routing paths and their corresponding commu-nication parameters, which jointly optimize the end-to-end delay and energy consumption for given QoS require-ments. In addition, the impact of the Internet on theachieved QoS is also taken into account. Moreover, we pro-pose a cross-layer communication protocol which canimplement the optimization framework in practical sce-narios. The results show that the proposed solution outper-forms existing layered solutions and the joint-objectivecross-layer solution can balance between different designobjectives.

The contributions of this paper are summarized asfollows:

� We provide an in-depth review of the state of the art inIoT-related initiatives and standardization efforts, tobetter motivate the need of a unified cross-layersolution.� We identify the interrelations among the device capa-

bilities, physical layer, link layer and network layer,and explain how these are captured in our solution.� We define a cross-layer optimization framework with a

single weighted joint-objective function according tothe service-dependent QoS requirements.

� We propose a cross-layer protocol to implement thecross-layer optimization framework in practicalscenarios.� We provide extensive simulation results which show

that our proposed solution outperforms existing layeredprotocols.

To the best of our knowledge, this is the first unifiedsolution for end-to-end efficient and reliable communica-tions in the IoT. The remainder of this article is organizedas follows. In Section 2, we revise the related work in termsof on-going standardization efforts for the IoT. In Section 3,we describe the reference network architecture for the IoTthat we consider throughout the paper. In Section 4, we de-scribe our design approach and develop the cross-layeroptimization framework which captures the existing rela-tions among the different layers of the protocol stack. InSection 5, we describe the cross-layer protocol operationneeded to implement our optimization framework in prac-tical scenarios. In Section 6, we evaluate the performanceof the proposed solution by means of simulation. Finally,we conclude the paper in Section 7.

2. Related work

Many research organizations are working towards thedeployment and standardization of the IoT. We summarizethe most relevant standardization efforts to date asfollows:

� IEEE 802.15.4 Standard [11]: provides the specificationsfor the physical layer (frequency spectrum allocation,modulation, data rates, and power control) and the linklayer (MAC and error control) for Low-Rate WirelessPersonal Area Networks (LR-WPANs). However, it doesnot specify the implementation of higher-layer func-tionalities (e.g., routing, end-to-end reliability) or howthe communication among things over the Internet isrealized.� IETF Low power Wireless Personal Area Networks (6LoW-

PAN) Standard [12]: defines a set of protocols to inte-grate low-complexity devices which operate under theIEEE 802.15.4 Standard into IPv6 networks. However,several challenges appear due to, among others, themismatch between the minimum packet size for IPv6networks and that of IEEE 802.15.4, or the difficulty tomanage routing tables for the expected number ofnodes involved in the IoT. Mechanisms to guaranteeend-to-end reliability are not provided, either.� IETF Routing Over Low power and Lossy networks (ROLL)

Working Group (WG) [13]: focuses on the developmentof new routing protocols for low power and lossy net-works, and appears as a complement to the 6LoWPAN.Unfortunately, there is still a long path for the WG toreach a complete solution. Moreover, its attention isonly on efficient routing and does not guarantee anyQoS requirement such as end-to-end delay or reliability.� ETSI Machine to Machine (M2M) Technical Commit-

tee [14]: concentrates mainly on ad hoc networksamong things in which Internet is not part of the sys-

Fig. 1. Abstract network architecture of the IoT.

624 C. Han et al. / Computer Networks 57 (2013) 622–633

tem. For the time being, no specifications are providedfor things’ addressing, location, and QoS. Furthermore,different devices use different network protocols tocommunicate within the same M2M network [15]. Thisdramatically increases the burden on the gateway,which needs to adapt every transmission among vari-ous devices.

For the time being, there are many independent solu-tions for the different layer functionalities in the protocolstack. In the following, we review some of the relatedwork, for instance:

� Physical layer: in [16], the authors advocate a physical-layer-driven approach to protocol design for wirelesssensor networks with emphasis on the underlying hard-ware parameters. In [17], the authors demonstrate theimportance of the physical layer modeling on the per-formance evaluation. Nevertheless, neither of themexploits the interrelation among the MAC and otherlayer functionalities with the physical layer and hence,their joint influence on the end-to-end communicationperformance.� Link layer: the authors in [18] present a broad overview

of the MAC protocols conducted in the field of wirelesssensor and ad hoc networks. However, they fail to pro-vide a vivid guidance on the choice of MAC techniquesand the associated parameters. By contrast, the authorsin [19] propose spatial correlation-based collaborativemedium access control (CC-MAC), an energy efficientMAC that exploits spatial correlation in wireless sensornetworks on the MAC layer. Both of them admit that theMAC layer plays an important role in the performanceof the overall system and affects other layers, whilethey ignore these effects and their impact on the systemperformance.� Network layer: the authors in [20] summarize the data

routing algorithms and classify the approaches intothree categories: data-centric, hierarchical and locationbased. In addition, in [21], the authors study the designtrade-offs between energy and communication over-head savings for the existing routing protocols. How-ever, neither of them is appropriate for IoT since thephysical attributes of nodes are not taken into account,which have direct impact on the performance and eventhe validity of the routing algorithms. Furthermore,they omit the interactions between routing algorithmand other layers.

As can be seen, these layered solutions cannot suc-cessfully capture the twofold heterogeneity of the IoT.In our vision, it is particularly challenging to develop a‘‘one-size-fits-all’’ solution by following a classical ap-proach. Alternatively, more advanced adaptive cross-layer schemes are necessary to guarantee efficient andreliable communication in the IoT. Cross-layer protocolshave been successfully developed in other paradigmssuch as WSNs [7,8], WMNs [9] or AHNs [10], among oth-ers. However, these cannot be directly used in the IoT,due to the major aforementioned differences in theparadigms.

3. Reference network architecture

We define next the reference network architecture forthe IoT which is considered throughout the paper. Weidentify the following network components:

� Things: i.e., physical objects with very diverse hardwarespecifications in terms of communication, computation,memory and data storage capacity, or transmissionpower. Personal electronic devices, home appliancesor all sorts of equipment, are examples of things.� Access Points (APs): i.e., more advanced devices which

play the role of local network coordinator as well asinterface and gateway for the communication over theInternet. We refer to the set of things under the controlof a single AP as the AP domain.� The Internet: i.e., a fundamental component of the IoT. In

our analysis, we treat the Internet as a black box whichis characterized by an stochastic queueing delay modeland a stochastic packet loss model [22].

Fig. 1 illustrates the network architecture of the IoT, inwhich several things are connected to the Internet via acommon AP. In common scenarios (e.g., at home, in the of-fice), the AP domain is composed by a few tens of things.Additional network considerations are summarized asfollows:

� Communication types: we distinguish two types of com-munication, namely, intra-AP and inter-AP. Intra-APcommunication is established among things withinthe same AP domain. Despite direct transmissionamong things might be possible, different capabilitiesamong them motivates the use of the AP for coordina-tion. Inter-AP communication is established amongthings in different AP domains. In this case, the APserves as a gateway as well as the coordinator.� Interconnection between things and AP: the AP is able to

directly communicate with all the things in its domainin a single hop. However, not all the things are able todirectly communicate with the AP, in which case

C. Han et al. / Computer Networks 57 (2013) 622–633 625

multi-hop links are required. This asymmetry in thelinks results from the fact that the transmission powerof the AP is generally higher than that of many things.� Network knowledge: the AP is able to collect information

about all the things in its domain (e.g., during the net-work association phase, as we explain in Section 5),such as device type, approximate location, communica-tion or computation capabilities, among others. By con-trast, things might only know the AP ID, but have noinformation about other things. Things are not able toperform complex computation locally.� Centralized decision: the AP is able to run optimization

algorithms locally by exploiting its network knowledge.Therefore, the computation complexity is shifted fromthings as in a distributed manner to the AP in a central-ized fashion. As a result, global optimal routes and com-munication parameter values for the AP domain can beobtained. This is feasible since the size of the AP domainis in the order of a few tens of things and, thus, theresulting complexity is affordable for standard AP hard-ware capabilities.

4. Cross-layer optimization framework

In this section, we present our new cross-layer optimi-zation framework for the IoT. First, we describe our ap-proach to define and design the optimization framework.Second, we describe the functionalities and interrelationsamong the physical layer, the link layer and the networklayer as well as the things’ capabilities. Third, we mathe-matically derive the framework.

4.1. Cross-layer design approach

We follow a resource allocation approach [23] to inte-grate different communication functionalities into onecoherent mathematical optimization model and to providean adaptive solution for cross-layer design and control. Inour framework, we incorporate multiple application-basedobjectives, with different scaling weights for different situ-ations, into one optimization problem. Additionally, oursolution relies in a centralized optimization model to jointlycontrol the parameters at the physical layer (channel,modulation), link layer (MAC, error control), and networklayer (addressing, routing) in the IoT, to ultimately reachthe optimality according to the application-dependentobjective function.

4.2. Multi-objective optimization

The IoT should provide differentiated services for appli-cations with different QoS requirements, ranging from er-ror-limited applications or minimum energyconsumption applications to highly-delay-sensitive appli-cations or any combination of them. Hence, we considera multi-objective optimization problem [24] which cansimultaneously optimize multiple conflicting objectivessubject to certain constraints, as follows:

min PERe2e; Ee2e; Te2ef g; ð1Þ

where PERe2e; Ee2e and Te2e stand for end-to-end packet er-ror rate, energy consumption and delay, respectively.

In order to solve this multi-objective optimization prob-lem, an intuitive approach is to construct a single aggregateobjective function which is defined by the weighted linearcombination of each objective. However, one concern thatarises in this situation is that the three objectives differ inthe units in which they are measured as well as their orderof magnitude. To resolve this, we normalize each term andoptimize their deviations with respect to a pre-definedthreshold, instead to minimize their absolute values. Theoverall objective function of the framework becomes:

min wPER �PERe2e

PERopt� 1

��������þwE �

Ee2e

Eopt� 1

��������wPER �

PERe2e

PERTH� 1

��������

�

þwT �Te2e

Topt� 1

���������; ð2Þ

where wPER þwE þwT ¼ 1 are the three linear weights forthe end-to-end packet error rate, energy consumptionand time delay objectives, and PERopt; Eopt; Topt are theend-to-end packet error rate, energy consumption and de-lay utopia values [25] for normalizing purposes, respec-tively. These utopia values are defined to be theunattainable minimum values, which are used to providethe non-dimensional objective functions and can be com-puted offline. Therefore, we have a single objective optimi-zation problem to solve. Accordingly to different QoSrequirements, we can control the emphasis of each termin the objective function (2) by assigning different weights.

With the objective of either minimizing the packet errorrate to improve the network efficiency, minimizing the en-ergy consumption to prolong the network lifetime, reduc-ing the network end-to-end delay to increase the systemthroughput, or any combination of these, the proposedcross-layer framework jointly selects the best combinationof transmit power and modulation (Section 4.3), the appro-priate error control mechanism and MAC parameter values(Section 4.4), and the proper addressing and routing algo-rithms (Section 4.5). We define next the functionalitiesimplemented at each layer, the tunable parameter valuesand the interrelations among layers, as well as the impactof the things capabilities in each of them.

4.3. Physical layer functionalities

At the physical layer, different things have differentmaximum transmission power, can select different modu-lation schemes, and have different data storage capacity(i.e., number of packets that thincs can locally queue).

4.3.1. Frequency allocation and channel modelIn our model, we consider that things follow the fre-

quency spectrum allocation defined by the IEEE 802.15.4standard [11]. Things are able to dynamically select amongany of the 5-MHz-wide sub-channels in the 2400–2480 MHz band. Since many of the applications of theIoT are indoor (e.g., home, office), we consider the ITUchannel model for indoor propagation [26]. The total pathloss L in dB is given by

626 C. Han et al. / Computer Networks 57 (2013) 622–633

L f ;dð Þ ¼ 20log10 fð Þ þ Nlog10 dð Þ þ Lf nð Þ � 28; ð3Þ

where f is the carrier frequency in MHz, d is the transmis-sion distance in meters, N is the distance attenuation coef-ficient (i.e., N = 20 in our simulations), and Lf is the floorpenetration loss factor and n is the number of floors be-tween the transmitter and the receiver (we consider onlyone floor in our current scenario, i.e., n ¼ 1; Lf ð1Þ ¼ 0).

4.3.2. Transmission power, modulation and bit error rateThe transmission power and the modulation have a di-

rect impact on the Bit Error Rate (BER). Over link i, The BER,BERi

link, is determined by the Signal-to-Noise Ratio (SNR),SNRi

link, and the modulation modi, as follows

BERilink f ;dð Þ ¼ W SNRi

link f ;dð Þ;modi� �

; ð4Þ

where W returns the BER for a given modulation and SNR,and it is well-known for standard modulations. The SNR oflink i; SNRi

link is given in dB by

SNRilink f ;dð Þ ¼ Pi

Tx � L f ;dð Þilink � Pnoise; ð5Þ

where PiTx is the transmission power in dB over link i and

Pnoise refers to the total noise power at the receiver in dB.We consider three standard modulations, namely, BPSK,

QPSK and 16-QAM, but any other modulation can be easilyincluded in the framework. These modulations are distin-guishable in terms of achievable BER for a given SNR, i.e.,W in (4), and spectral efficiency, i.e., theoretical achievabledata bit-rate for a given transmission bandwidth. A highercomplexity modulation exhibits higher bandwidth effi-ciency, which results in a higher transmission data rateor a shorter transmission time, Ti

data. However, these comewith the cost of an increase in the energy consumption atthe transmitter and the receiver due to the increase in thecomputation complexity, as well as, potentially, also in theprocessing time. Furthermore, more complex modulationsrequire a higher SNRi

link, thus, higher PiTx, to achieve the

same BERilink. At the same time, though, the transmission

time, Tidata, is shorter, which also affects the link energy

consumption. These interrelations are properly capturedin the framework.

4.3.3. Data storage capacity and packet dropout probabilityThe data storage capacity mem of the things affects the

packet dropout rate, i.e., the probability of discarding apacket at link i; Pi

packet-dropout , due to the fact that it cannotbe queued at the transmitter or at the receiver. This is gi-ven by

Pipacket-dropout ¼ C memi;Ri

traffic

� �; ð6Þ

where C is a function that relates the maximum number ofpackets memi that can be queued at the transmitter or thereceiver and the total local traffic (own traffic and relayedtraffic), Ri

traffic . For example, in the simplest case, Poissontraffic can be assumed, and transmitter and receiver canbe modeled as a single server queue with a buffer of sizememi.

4.4. Link layer

At the link layer, we analyze the impact of the errorcontrol mechanism and the MAC protocol on the networkperformance, as well as, their interrelations with other lay-ers and the limitations imposed by the things capabilities.

4.4.1. Error control and packet error rateAmong the options for error control, Forward Error Cor-

rection (FEC) codes are used to fix erroneous bits intro-duced by the channel, the noise and the interference, andultimately to reduce the Packet Error Rate of linki;PERi

link. Specifically, if we consider the use of block codes,for a (n, k, t) code with block size n and k bits of informa-tion, t errors per block can be successfully corrected atmost. In our framework, we use the following notationsfor n ¼ Nbits and k ¼ Ndata

bits . Despite FEC codes improve thePER, the addition of NFEC

bits redundant bits reduces the effec-tive transmission data rate by Ndata

bits =Nbits.Another common error control mechanism is Automatic

Repeat reQuest (ARQ), which is based on the retransmis-sion of the whole packet in case of erroneous reception.This scheme is spectrally efficient only when the channelconditions are favorable, i.e., SNRi

link is high, or BERilink is

low, since no redundancy bits are introduced in the packet.If the BERi

link is relatively high, the number of retransmis-sions increases, and so do both the energy consumptionand delay.

To exploit the benefits from FEC codes for poor qualitychannel conditions, i.e., when the SNRi

link is low, as wellas the merits of ARQ when the channel conditions aregood, i.e., when the SNRi

link is high, we advocate for theuse of a Hybrid ARQ scheme [27]. Hybrid ARQ schemes re-sult from the combination of the two approaches. Initially,an uncoded or lightly coded packet is transmitted. If the re-ceived packet has more errors than those that can be cor-rected by the chosen FEC code, a more robust FEC code ischosen.

In our framework, things make use of Bose–Chaudh-uri–Hocquenghem (BCH) codes [28]. BCH codes are usedinstead of more complex codes such as ConvolutionalCodes, because BCH codes have higher energy efficiencyand lower computation complexity [29]. These two fac-tors are critical for things with very limited energy stor-age and computation capability. In transmission, theinitial packet is either not coded or coded with a (128,106, 3) BCH code to reduce PERi

link without drastically sac-rificing the transmission data rate. If the first transmissionfails, i.e., the number of errors is larger than that can becorrected, a more robust FEC code is used for the retrans-mitted packet (e.g., (128, 78, 7)) until the packet is suc-cessfully decoded or the maximum number oftransmissions is expired. Using this Hybrid ARQ errorcontrol scheme, the overall packet error rate over link iis given by

PERilink-RTR ¼ � ðPERi

link-uncoded;Nilink-max;N

FECbitsÞ; ð7Þ

where � is a function that relates the PER of link i after Hy-brid ARQ error control, PERi

link-RTR, with the uncoded linkPER PERi

link-uncoded. In the above equation, NFECbits is the FEC

C. Han et al. / Computer Networks 57 (2013) 622–633 627

redundancy length and Nilink-max is the maximum number of

transmissions including retransmissions, which are bothadjustable in our cross-layer framework.

As can be seen, larger NFECbits and larger Ni

link-max lead to (i) alower transmission data rate or larger link delay, (ii) a low-er PER, and (iii) a higher total energy consumption. The im-pact of the error control parameters interplays with thosemade by adjusting the transmission power and the modu-lation scheme at the physical layer. Hence, it is beneficialto explore the interactions among the physical layer andlink layer functionalities and jointly determine the trans-mission power, the modulation scheme, and the error con-trol parameters.

4.4.2. Medium access controlWe consider a variation of Sleep MAC (SMAC) [30] and

Carrier Sense Multiple Access with Collision Avoidance(CSMA/CA), for mainly two reasons. First, idle listening isa major source of energy consumption in low traffic appli-cations such as those expected in the IoT, and many of thephysical objects in the IoT may have very limited energystorage. Therefore, we adopt the idea of SMAC in whichthings periodically listen and sleep [30]. For example, al-most, 90% of the total energy consumed during idle listen-ing can be saved when things use an awake/sleep dutycycle of 10%. However, by decreasing the duty cycle, thechances for things to be connected decrease too and thelink delay increases. Second, the interference among thethings in one AP domain or in close AP domains affectsthe SNR and thus degrades the BER and the PER. In theCSMA/CA, a node attempts to reserve the channel after itsees the medium idle for an Inter-Frame Space (IFS)amount of time. If the node fails to reserve the medium,it switches to sleep mode to save energy and waits forthe next listening cycle. This medium access method caneliminate the interference drastically if the carrier sensingis properly performed. As a result, the hybrid of SMAC andCSMA/CA medium access protocol can save the energy aswell as reduce the interference among the things.

In our framework, the duration of the listen and sleepcycles (Tlisten; Tsleep ¼ 9� Tlisten for a 10% duty cycle) areadaptive to the QoS requirements and they are set thesame for all nodes in one AP domain. The longer the sleepduration is, the lower the idle energy consumption, but thelonger end-to-end delay. This duration parameter in theMAC protocol is taken into account in our cross-layerframework to interplay with the physical layer parametersas well as Hybrid ARQ parameters.

4.5. Network layer

At the network layer, we look at the things addressingand information routing in the IoT. In addition, we discussthe impact of the packet size on the different layers.

4.5.1. Addressing of thingsThe IoT is expected to have an incredibly high number

of things and each of them should be retrievable with a un-ique address. Consistently with 6LoWPAN, we advocate forthe use of IPv6 addressing for the IoT. IPv6 addresses areexpressed by 128 bits, which allow the definition of 1038

unique addresses (these are expectedly enough for thetime being). However, IPv6 addresses are only used for in-ter-AP communications, while much shorter local ad-dresses are used in intra-AP communications. The APreplaces the thing local address by the thing IPv6 full ad-dress for communication with other AP domains.

4.5.2. RoutingWe consider a destination-based routing mechanism

[21] where the AP selects the end-to-end route and theconfiguration parameters for each link in a centralizedmanner, due to the following reasons. First, it is consis-tent with the Internet hierarchical architecture composedof the Internet, gateway, router and end users from top tobottom. Second, only the AP has information of the thingsin its operation region and is capable to perform optimi-zation computations. Third, the complexity of this cen-tralized routing is comparable to that of existingdistributed cross-layer solutions. By contrast, our frame-work achieves the global optimal solution, contrary tothe distributed case, in which the end-to-end optimalproblem is solved by doing only local optimization linkby link.

4.5.3. Impact of the packet lengthIn our framework, a fixed packet size Nbits is selected

and used for all the links throughout a given path. A largerpacket size results in a reduced end-to-end delay by savingthe handshake time Ti

handshake, the acknowledgement timeTi

ack, and the queuing time Tiqueueing that are required for

each packet. Additionally, the reduction of the total num-ber of packets to be sent has an impact on the total energyconsumption, while at the same time, transmitting morebits of information in a packet affects the PER. All theseinterrelations are incorporated in our cross-layerframework.

So far, we have explored the interrelations among theparameters at the physical layer (Section 4.3), the linklayer (Section 4.4) and the network layer 4.5. These param-eters including the transmission power, modulation type,the FEC length, the number of retransmissions, the listenduration, the packet size and their interactions, are all cap-tured in our cross-layer framework, as described in the fol-lowing section.

4.6. Mathematical framework

By starting from the knowledge of the things in its do-main, the AP runs the optimization algorithm locally, find-ing the optimal domain path and the correspondingcommunication parameters, which include transmissionpower at link i; Pi

link, modulation type, modi, packet size,Nbits, additional bits due to FEC codes NFEC

bits , the maximumnumber of transmissions including retransmissionNi

link-max, and listen time duration Tlisten. The cross-layeroptimization problem is mathematically defined asfollows:

Given ðofflineÞ : PERopt; Eopt; Topt; PERTH; ETH; EkTH; TTH;RTH;

Nheaderbits ; PERInternet; TInternet; ð8Þ

628 C. Han et al. / Computer Networks 57 (2013) 622–633

Cð�Þ; � ð�Þ;Wð�Þ; c; Tihandshake; T

idata; T

itimeout ; T

iack; T

iqueueing ;memi:

ð9Þ

Compute ðofflineÞ : wPER;wE;wT ;Nilink;R

itraffic: ð10Þ

Find : PiTx;modi

;Nbits;NFECbits ; Tlisten;N

ilink�max: ð11Þ

Minimize : wPER �PERe2e

PERopt� 1

��������þwE �

Ee2e

Eopt� 1

��������þwT

� Te2e

Topt� 1

�������� ð12Þ

Subject to : wPER þwE þwT ¼ 1; ð13Þ

PERe2e ¼ 1� ð1� PERmulti�hopsÞð1� PERInternetÞ� PERTH; ð14Þ

Nbits ¼ Nheaderbits þ NFEC

bits þ Ndatabits ; ð15Þ

PERilink�RTR ¼ � ðPERi

link�uncoded;Nilink;N

FECbitsÞ; ð16Þ

Nilink ¼ ð1� PERi

link�uncodedÞ�1; ð17Þ

Ek ¼ Nbits � Ekb � Ek

TH; ð18Þ

Ee2e ¼XNhop

k¼1

Ek � ETH; ð19Þ

Ti � ðTihandshake þ Ti

data þ TitimeoutÞ � ðN

ilink � 1Þ

þ ðTihandshake þ Ti

data þ TiackÞ þ Tsleep þ Ti

DSP; ð20Þ

Te2e ¼XNhop

i¼1

ðTiqueueing þ TiÞ þ TInternet; ð21Þ

PðTe2e � TTHÞ � c; ð22Þ

PðTe2e � TTHÞ

�var TqueueingþInternet� �

var TqueueingþInternet� �

þ TTH�PNhop

i¼1 Ti�TqueueingþInternet

� �2

�1�c;ð23Þ

Re2e ¼Ndata

bits

Te2e� RTH: ð24Þ

The notation in the framework is as follows:

� Restricted by the threshold PERTH;PERe2e is the end-to-end packet error rate, which is dependent on the PERof the multi-hop transmission, PERmulti-hops, and the PERof the Internet, PERInternet .� PERmulti-hops is a function of the PER over link i with

Hybrid ARQ error control, PERilink-RTR, given by

PERmulti-hops ¼ 1�YNhop

i¼1

1� PERilink-RTR

� �: ð25Þ

� Pipacket-dropout is the packet dropout rate over link i, and

BERilink is the bit error rate over link i, which is a function

of signal to noise ratio over the link SNRilink and the

modulation type modi (see (4)). These parametersdetermine the uncoded packet error rate over link i asfollows:

PERilink-uncoded ¼ 1� Pi

packet-dropout

� �

� 1� 1� BERilink

� �Nbits

: ð26Þ

� Nbits is the packet size, which contains the headerlength, data length and the FEC redundancy length.� Ni

link is the upper-bound for the number of transmis-sions of a packet with correctly decoding, over link i.� Ek is the energy on kth node, which is the product of the

packet size Nbits and the energy required for one bit, Ekb.

� Ekb is calculated as

Ekb ¼ 2Eelec

b þ PkTx

Rktraffic

; ð27Þ

where Eelecb ¼ Eelec

b�Tx ¼ Eelecb�Rx in Joule/bit is the distance-inde-

pendent energy to transmit one bit. Eelecb�Tx is the energy per

bit needed by the transmitter electronics (PLLs, VCOs,AMPs, DSP, etc.) and Eelec

b�Rx is the energy per bit utilizedby the receiver electronics.� Ee2e is the overall energy consumption over the entire

path, with the constraint ETH .� Ti is the delay at link i excluding the queueing delay. It

is composed of the time for handshake Tihandshake, time for

data transmission Tidata, timeout delay Ti

timeout , time foracknowledgement Ti

ack, sleep time Tsleep, and signal pro-cessing time Ti

DSP .� Restricted by the constraint TTH; Te2e is the end-to-end

time duration including the Internet delay TInternet forinter-AP communications, and the link queueing delayTi

queueing . Both of the queueing delay and the Internetduration are determined by many factors like currenttraffic, other nodes behavior or hardware status, amongothers. By Central Limit Theorem, we can model theoverall end-to-to delay Te2e as a Gaussian distributedrandom variable with mean

PNhop

i¼1 Ti þ TqueueingþInternet

� �and variance var TqueueingþInternet

� �� �, where

TqueueingþInternet ¼PNhop

i¼1 Tiqueueing þ TInternet .

� For a target probability c, we use the Chebyshev’sinequality [31] to decompose the end-to-end delay con-straint into PðTe2e P TTHÞ 6 1� c, by satisfying

TTH �XNhop

i¼1

Ti � TqueueingþInternet > 0: ð28Þ

� Re2e is the end-to-end throughput, which is a inverselyproportional to the end-to-end delay Te2e.� Table 1 summarizes the parameters that consist of the

optimization solution.

Table 1Adaptive parameters in the cross-layer framework.

Layer Parameters

Physical layer Transmission power PiTx

Modulation scheme modi

Link layer FEC coding scheme

Coding length NFECbits

Maximum number of transmissions Nilink�max

Listening period length Tlisten

Network layer Routing pathPacket length Nbits

Fig. 2. Illustration of the optimization problem in a directed acyclicgraph.

C. Han et al. / Computer Networks 57 (2013) 622–633 629

The proposed optimization problem can be solved in twosteps.

� Formulation of the directed acyclic graph. According tothe objective function, the constraints, and the hard-ware capabilities of things, the optimization problemcan be expressed as a corresponding shortest-pathproblem in a directed acyclic graph, e.g., Fig. 2 for anintra-AP communication scenario. Node S is the source,node D is the destination, and A, B, C, E, F, G, H, I, J areintermediate nodes. Since the nodes have distinct com-munication abilities, the coverage range, denoted by thedashed circle (for a clear view, only the coverage rangecircles for node A and H are shown), varies for eachnode. The costs associated to the directional link fromnode i to j, Cij, are derived accordingly to be consistentwith the objective function. Without lose of generality,this graph can be extended for the inter-AP communica-tion, by changing node D to be the AP for the Internetconnection. In addition, a bilateral Internet linkbetween the two APs and the topology for the destina-tion AP range need to be added.� Solving the shortest path problem. So far, the optimiza-

tion problem has been transformed into a shortest pathproblem in the directed acyclic graph, with nodes,

directional edges, different coverage, and associatedcosts. The goal of this shortest path problem is to selectthe path with the smallest cost from S to D. Many algo-rithms can solve this problem, e.g., Dijkstra’s algorithm[32], which functions by constructing a shortest-pathtree from the source node to every other node in thegraph, and stops once the shortest path to the destina-tion node has been determined. The complexity of Dijk-stra’s algorithm is bounded by OðV2Þ, where V is thenumber of nodes involved.

5. Protocol operation

In this section, we propose a cross-layer protocol whichimplements our cross-layer optimization framework inpractical scenarios. We divide the protocol operation infour phases:

� Phase I: Service discovery & Network association. The APperiodically broadcasts its ID and supported communi-cation parameters within its domain. This is feasiblesince we consider that the AP has sufficiently largepower to directly communicate with every thing in itsdomain (Section 3). Correspondingly, the things in theAP domain register themselves to the AP by means ofsending a Network Association (NAS) packet. This regis-tration can be done either with direct communicationor in a multi-hop manner, depending on the locationand the transmission power of each thing. We cannotassume that things can have a notion of their location.However, the AP can cyclically modify its transmissionpower and define virtual regions. Only those things ina closer virtual region than the current transmitter areeligible to forward the network NAS packet towardsthe AP. By following this procedure, the AP collectsinformation about all the things in its domain, e.g.,device type, communication capabilities and logicallocation. Things can update their network associationat any time by sending a new NAS packet. NAS packetsare transmitted by following the normal MAC operationdescribed in Section 4.4.� Phase II: Transmission initiation. When a node has data to

transmit, it first checks if it has a valid route for the des-tination (some things may be able to keep a local rout-ing table, some others might not). If it has a timelyroute, it proceeds with the transmission by followingthe MAC normal operation described in Section 4.4.Otherwise, it generates a Route Request (RR) packetcontaining the destination thing ID. The RR packet istransmitted towards the AP by following the normalMAC operation. When an intermediate node receives aRR packet, it checks its current available resources andits logical location. If it has sufficient resources and itis closer to the AP, it forwards the RR packet afterappending its updated information to it. Otherwise, itignores the packet and continues with its normaloperation.� Phase III: Route definition. The AP receives the RR packet

over several paths, whose intermediate nodes are listedas priority candidates for data transmission (the RRpacket contains the most updated information about

Table 2Random Variables (RVs) for the parameters in the cross-layer framework.

RV Distribution

Packet error rate of theInternet

PERInternet � U 0;10�4� �

Packet dropout rate overlink i

Pipacket�dropout � U 0;0:1ð Þ

Internet delay TInternet �N 102;104� �

ms

Things queuing delay ateach link

Tiqueueing �N 10;104

� �ms

Noise at each link noisei �N 0; Pnoiseð Þ, where10log10Pnoise ¼ �86 dB

630 C. Han et al. / Computer Networks 57 (2013) 622–633

some of the things in the domain). Then the AP formu-lates the optimization framework given the potentialpath and QoS requirements to find the optimal pathand the associated communication parameters, asexplained in Section 4.6. The decision depends on theQoS requirements, the location of the source and desti-nation things, and whether it is an intra-AP (in whichthe AP may or may not participate in the transmission)or inter-AP transmission (in this case, the AP broadcastthe optimal route from the source thing to itself). Afterfinding the optimal path and communication parame-ters, the AP directly informs the chosen nodes with asingle broadcast transmission.� Phase IV: Message Transmission. Upon the reception of

the route information directly from the AP, the destina-tion thing sends a Route Acknowledge (RA) packet tothe previous hop in the route in order to acknowledgeits availability. The process is repeated hop-by-hopuntil the RA packet reaches the source. After this, thedata is transmitted by following the optimal route withthe chosen communication parameters and accordingto the MAC normal operation described in Section 4.4.If the source thing does not receive the RA packet aftera given time-out, it sends a new RR packet to the AP,with a flag to mark that the previous route was notvalid. The AP can decide to compute a new route orforce Phase I.

In this AP-oriented architecture, the computation com-plexity is shifted from things to the AP. Furthermore, theAP coordination can effectively diminish the impact of hav-ing the multiple-flow problems, and provide a globallyoptimal solution.

6. Performance evaluation

In this section we compare the performance achievedby our cross-layer solution against that achieved by tradi-tional layered solutions, in which individual communica-tion functionalities operate in isolation. We compare theresults when the QoS is focused on either end-to-end delayminimization, energy consumption minimization, or a lin-eal combination of both, while the PER is constraint to bebelow PERTH ¼ 10�4.

In our simulations, the total amount of data to transmitper transmission is 105 bits and the possible packet sizesNbits are 200, 500 or 2000 bits. For each transmission, thething randomly selects its destination. On average, 50% ofthe links are inter-AP and 50% of the links are intra-AP.Things have diverse hardware capabilities, in terms ofcomputing, memory, energy storage, power and communi-cation. Specifically, the link rate of things ranges among[250, 103, 104] kbps and the power of things varies among[10, 30, 50, 80, 100] milliwatts. Besides these deterministicparameter values, we summarize in Table 2 the randomvariables that are considered in our simulations. The errorbars in the figures represent the uncertainty interval at the95% confidence level. The four layered solutions that areplotted for comparison differ in the modulation scheme

and the packet size, and make use of the shortest pathrouting.

Fig. 3a and b show the end-to-end delay in millisecondsand the energy consumption in microjoules versus thenumber of things in the IoT network. The distance betweenthe transmitter and the receiver is 40 m. Our cross-layersolution has at least 10% gain over other layered solutions.In addition, we can observe that neither the end-to-end de-lay nor the energy consumption increases as the number ofthings increases. This can be explained that the highernode density essentially create more options of paths fortransmission, while the end-to-end consumption has noproportional relationship with it. The 95% confidence inter-val imply that our cross-layer solutions have better stabil-ity than other layered solutions, and the end-to-endperformance does not necessarily change although thecomputation complexity at the AP increases.

The end-to-end delay and the energy consumptioncurves are shown in Fig. 4a and 4b as a function of thedistance between source and destination. Since many ofthe applications of IoT are indoor (e.g., home, office), weconsider the distance less than 40 m. In our simulation,the distance increases from 10 m to 40 m and totally 10things are involved. When the transmission distance is in-creased, the propagation delay increases and the numberof nodes in the path may increase. We can neglect theinfluence of the propagation delay over this order of dis-tance. However, increasing the number of hops in the pathimplies the rise of the end-to-end delay since there areadditional handshake, processing and queuing delay intro-duced into the transmission. We can observe the trend ofthe increase in Te2e as the distance grows. Similarly, the en-ergy consumed for the longer distance and by the addi-tional nodes increase the overall Ee2e. As shown in Fig. 4,the performance gain of our cross-layer solution over thebest mod=Nbits combination increases with the distance,which further proves the advantage of our cross-layer solu-tion over the layered solutions. Among the four layeredsolutions considered, the one with the simplest modula-tion scheme and the largest packet size performs the best,although it has around 10% performance degradation com-pared to our cross-layer solution.

Moreover, we investigate the performance differencebetween the cross-layer solution with single objectiveand that with joint objectives, in Fig. 5a and 5b. InFig. 5a, we compare the Te2e performance among cross-layer solutions with a single-objective (minimizing either

10 15 20 25 30 35 400

2000

4000

6000

8000

10000

12000

14000

Number of things in the IoT network

End−

to−e

nd d

elay

in m

illise

cond

sEnd−to−end delay versus number of things

BPSK maxPacketBPSK minPacket16QAM maxPacket16QAM minPacketCross−layer

(a) Delay vs. number of things

10 15 20 25 30 35 40

0

2000

4000

6000

8000

Number of things in the network

Ener

gy c

onsu

mpt

ion

in m

icro

joul

es

Energy consumption versus number of things

BPSK maxPacketBPSK minPacket16QAM maxPacket16QAM minPacketCross−layer

(b) Energy consumption vs. number of things

Fig. 3. Comparison between layered solutions and cross-layer solutions with respect to the number of things.

10 15 20 25 30 35 400

2000

4000

6000

8000

10000

Distance in meters

End−

to−e

nd d

elay

in m

illise

cond

s End−to−end delay versus distance

BPSK maxPacketBPSK minPacket16QAM maxPacket16QAM minPacketCross−layer

(a) Delay vs. distance

10 15 20 25 30 35 400

1000

2000

3000

4000

Distance in meters

Ener

gy c

onsu

mpt

ion

in m

icro

joul

es

Energy consumption versus distance

BPSK maxPacketBPSK minPacket16QAM maxPacket16QAM minPacketCross−layer

(b) Energy consumption vs. distance

Fig. 4. Comparison between layered solutions and cross-layer solutions with respect to the distance.

10 15 20 25 30 35 400

500

1000

1500

2000

Distance in meters

End−

to−e

nd d

elay

in m

illise

cond

s End−to−end delay versus distance

minEnergyminDelaymin(Delay+Energy)

(a) Delay vs. distance

10 15 20 25 30 35 40−200

0

200

400

600

800

Distance in meters

Ener

gy c

onsu

mpt

ion

in m

icro

joul

es

Energy consumption versus distance

minEnergyminDelaymin(Delay+Energy)

(b) Energy consumption vs. distance

Fig. 5. Comparison between single-objective cross-layer solutions and joint-objective cross-layer solution with respect to the distance.

C. Han et al. / Computer Networks 57 (2013) 622–633 631

Te2e or Ee2e) and cross-layer solution with a joint-objective(minimizing f0:5 � Te2e þ 0:5 � Ee2eg). The cross-layer solu-tion with single objective of minimizing Te2e has the small-est end-to-end delay with an increasing performance gainover the other two solutions. Similar to the energy con-sumption curves shown in Fig. 5b, the single objectivesolution outperforms for its focused criterion, while it

exhibits degradation for the other. For example, the solu-tion with the objective of minimizing energy consumptionconsumes the least amount of energy, but it introduces thelargest end-to-end delay. On the contrary, the cross-layersolution with joint objectives of minimizing Te2e as wellas Ee2e achieves satisfactory performance in both end-to-end delay and energy consumption.

632 C. Han et al. / Computer Networks 57 (2013) 622–633

7. Conclusion

The IoT is the enabling technology for a plethora oflong-awaited applications and business opportunities inthe fields of domotics, e-health, real-time monitoringand logistics, among others, by allowing the seamlesscommunication among all sorts of physical devices. Thereare many on-going world-wide research initiatives andstandardization efforts, which aim at making the IoT areality. However, the very high heterogeneity in hard-ware capabilities of things and QoS requirements for dif-ferent applications limits the performance of classicallayered protocol solutions and the existing cross-layersolutions for wireless sensor networks or ad hocnetworks.

In this paper, we have explored the interaction amongfunctionalities across the different layers in the protocolstack and developed a novel cross-layer optimizationframework for the IoT. This framework captures theseinterrelations among different layers, the twofold hetero-geneity of things and the impact of the Internet in the net-work performance. Moreover, we have proposed a cross-layer protocol to practically implement the proposed opti-mization framework. The results show that the proposedsolutions outperforms existing layered solutions byexploiting the interactions among layers and the implicitlycentralized network architecture of the IoT.

Acknowledgment

This work was supported by la Fundación Caja Madrid.

References

[1] D. Miorandi, S. Sicari, F. Pellegrini, I. Chlamtac, Internet of things:vision applications & research challenges, Ad Hoc Networks(Elsevier) Journal (2012).

[2] M. Weiser, The computer for the 21st century, Scientific American265 (3) (1991) 94–104.

[3] ITU Internet Reports, The Internet of Things, InternationalTelecommunication Union, Tech. Rep., 2005.

[4] L. Atzori, A. Iera, G. Morabito, The internet of things: a survey,Computer Networks (Elsevier) Journal 54 (2010) 2787–2805.

[5] S. Gusmeroli, S. Haller, M. Harrison, K. Kalaboukas, M. Tomasella, O.Vermesan, H. Vogt, K. Wouters, Vision and challenges for realisingthe internet of things, in: P. Friess, P. Guillemin, H. Sundmaeker, S.Woelfflé, (Eds.), European Commission, 2010.

[6] I.F. Akyildiz, J.M. Jornet, The internet of nano-things, IEEE WirelessCommunications Magazine 17 (6) (2010) 58–63.

[7] M.C. Vuran, I.F. Akyildiz, Xlp: a cross-layer protocol for efficientcommunication in wireless sensor networks, IEEE Transactions onMobile Computing 9 (11) (2010) 1578–1591.

[8] D. Pompili, I. Akyildiz, A multimedia cross-layer protocol forunderwater acoustic sensor networks, IEEE Transactions onWireless Communications 9 (9) (2010) 2924–2933.

[9] I. Akyildiz, X. Wang, Cross-layer design in wireless mesh networks,IEEE Transactions on Vehicular Technology 57 (2) (2008) 1061–1076.

[10] M. Conti, G. Maselli, G. Turi, S. Giordano, Cross-layering in mobile adhoc network design, Computer 37 (2) (2004) 48–51.

[11] IEEE 802.15 Wireless Personal Area Networks Task Group 4. <http://www.ieee802.org/15/pub/TG4.html>.

[12] N. Kushalnagar, C. Schumacher, G. Montenegro, IPv6 Over Low-Power Wireless Personal Area Networks (6LoWPANs): Overview,Assumptions, Problem Statement, and Goals, IETF, 2007.

[13] IETF Routing Over Low power and Lossy networks (ROLL). < https://datatracker.ietf.org/doc/charter-ietf-roll/>.

[14] Z. Shelby, ETSI M2M Standardization, 2009.[15] Z. Fadlullah, M. Fouda, N. Kato, A. Takeuchi, N. Iwasaki, Y. Nozaki,

Toward intelligent machine-to-machine communications insmart grid, IEEE Communications Magazine 49 (4) (2011)60–65.

[16] E. Shih, S. Cho, N. Ickes, R. Min, A. Sinha, A. Wang, A. Chandrakasan,Physical layer driven protocol and algorithm design for energy-efficient wireless sensor networks, in: Proceedings of the 7th ACMAnnual International Conference on Mobile Computing andNetworking, 2001, pp. 272–287.

[17] M. Takai, J. Martin, R. Bagrodia, Effects of wireless physical layermodeling in mobile ad hoc networks, in: Proceedings of the 2ndACM International Symposium on Mobile Ad Hoc Networking &Computing, 2001, pp. 87–94.

[18] S. Kumar, V. Raghavan, J. Deng, Medium access control protocols forad hoc wireless networks: a survey, Ad Hoc Networks (Elsevier)Journal 4 (3) (2006) 326–358.

[19] M. Vuran, I. Akyildiz, Spatial correlation-based collaborativemedium access control in wireless sensor networks, IEEE/ACMTransactions on Networking 14 (2) (2006) 316–329.

[20] K. Akkaya, M. Younis, A survey on routing protocols for wirelesssensor networks, Ad Hoc Networks (Elsevier) Journal 3 (3) (2005)325–349.

[21] J. Al-Karaki, A. Kamal, Routing techniques in wireless sensornetworks: a survey, IEEE Transactions on WirelessCommunications 11 (6) (2004) 6–28.

[22] W. Jiang, H. Schulzrinne, Modeling of packet loss and delay and theireffect on real-time multimedia service quality, in: Proceedings ofNOSSDAV, Citeseer, 2000.

[23] X. Lin, N. Shroff, R. Srikant, A tutorial on cross-layer optimization inwireless networks, IEEE Journal on Selected Areas inCommunications 24 (8) (2006) 1452–1463.

[24] K. Deb, Multi-objective optimization, Search Methodologies (2005)273–316.

[25] R. Marler, J. Arora, Survey of Multi-Objective Optimization Methodsfor Engineering, vol. 26, Springer, 2004 (no. 6).

[26] J. Seybold, Introduction to RF Propagation, Wiley Online, Library,2005.

[27] M. Vuran, I. Akyildiz, Error control in wireless sensor networks: across layer analysis, IEEE/ACM Transactions on Networking 17 (4)(2009) 1186–1199.

[28] S. Wicker, Error Control Systems for Digital Communication andStorage, Prentice-Hall Inc., 1994.

[29] Y. Sankarasubramaniam, I. Akyildiz, S. McLaughlin, Energy efficiencybased packet size optimization in wireless sensor networks, in:Proceedings of the First IEEE International Workshop on SensorNetwork Protocols and Applications, 2003, pp. 1–8.

[30] W. Ye, J. Heidemann, D. Estrin, Medium access controlwith coordinated adaptive sleeping for wireless sensornetworks, IEEE/ACM Transactions on Networking 12 (3) (2004)493–506.

[31] P. Huber, The behavior of maximum likelihood estimates undernonstandard conditions, in: Proceedings of the Fifth BerkeleySymposium on Mathematical Statistics and Probability, vol. 1(1),1967, pp. 221–233.

[32] D. Johnson, A note on Dijkstra’s shortest path algorithm, Journal ofthe ACM (JACM) 20 (3) (1973) 385–388.

Chong Han received the Bachelor of Engi-neering in Electrical Engineering and Tele-communications from The University of NewSouth Wales, Sydney, Australia, in 2011. Cur-rently, he is a graduate research assistant inthe Broadband Wireless Networking Labora-tory (BWN Lab), School of Electrical andComputer Engineering, Georgia Institute ofTechnology, Atlanta, GA. He is pursuing hisPh.D. degree under the supervision of Prof. IanF. Akyildiz. He is a student member of theIEEE. His current research interests are in the

Internet of Things, Internet of Nano-Things, Electromagnetic Nanonet-works, Graphene-enabled Wireless Communication and Terahertz BandCommunication Networks.

C. Han et al. / Computer Networks 57 (2013) 622–633 633

Josep Miquel Jornet received the EngineeringDegree in Telecommunication Engineeringand the Master of Science in Information andCommunication Technologies from the Schoolof Electrical Engineering, Universitat Politèc-nica de Catalunya (UPC), Barcelona, Spain, in2008. From September 2007 to December2008, he was a visiting researcher at theMassachusetts Institute of Technology, Cam-bridge, under the MIT Sea Grant program.Currently, he is pursuing his Ph.D. degree inthe Broadband Wireless Networking Labora-

tory (BWN Lab), School of Electrical and Computer Engineering, GeorgiaInstitute of Technology, Atlanta, with a fellowship from ’’la Caixa’’ (2009–2010) and Fundación Caja Madrid (2011–2012). He is a student member

of the IEEE and the ACM. His current research interests are in electro-magnetic nanonetworks, graphene-enabled wireless communication,Terahertz Band communication networks and the Internet of Nano-Things.

Etimad A. Fadel received the Bachelors degree in Computer Science atKing Abdul Aziz University with Senior Project title ATARES: ArabicCharacter Analysis and Recognition in 1994. She was awarded the Mphil/PhD degree in computer science at De Montfort University (DMU) withThesis title Distributed Systems Management Service in 2007. Currentlyshe is working as Assistant Professor at the Computer Science Departmentat KAU. Her main research interest is Distributed Systems, which aredeveloped based on middleware technology. Currently she is looking intoand working on Wireless Networks, Internet of Things and Internet ofNano-Things.

Ian F. Akyildiz received the B.S., M.S., andPh.D. degrees in Computer Engineering fromthe University of Erlangen-Nürnberg, Ger-many, in 1978, 1981 and 1984, respectively.Currently, he is the Ken Byers Chair Professorin Telecommunications with the School ofElectrical and Computer Engineering, GeorgiaInstitute of Technology, Atlanta, the Directorof the Broadband Wireless Networking Labo-ratory and Chair of the TelecommunicationGroup at Georgia Tech. Dr. Akyildiz is anhonorary professor with the School of Elec-

trical Engineering at Universitat Politècnica de Catalunya (UPC) in Bar-celona, Catalunya, Spain and the founder of N3Cat (NaNoNetworkingCenter in Catalunya). He is also an honorary professor with the Depart-

ment of Electrical, Electronic and Computer Engineering at the Universityof Pretoria, South Africa and the founder of the Advanced Sensor Net-works Lab. Since 2011, he is a Consulting Chair Professor at the Depart-ment of Information Technology, King Abdulaziz University (KAU) inJeddah, Saudi Arabia. Since September 2012, Dr. Akyildiz is also a FiDiProProfessor (Finland Distinguished Professor Program (FiDiPro) supportedby the Academy of Finland) at Tampere University of Technology,Department of Communications Engineering, Finland. He is the Editor-in-Chief of Computer Networks (Elsevier) Journal, and the founding Editor-in-Chief of the Ad Hoc Networks (Elsevier) Journal, the Physical Com-munication (Elsevier) Journal and the Nano Communication Networks(Elsevier) Journal. He is an IEEE Fellow (1996) and an ACM Fellow (1997).He received numerous awards from IEEE and ACM. His current researchinterests are in nanonetworks, Long Term Evolution (LTE) advancednetworks, cognitive radio networks and wireless sensor networks.